On Alan Turing's anticipation of connectionism (Q1293032): Difference between revisions

F. Rosenblatt used the term ``connectionist'' in the late 1950s to denote the view that information is stored in nervous systems in the form of new connections or transmission channels. The authors describe the 1948 National Physical Laboratory report by Alan Turing, ``Intelligent machinery'' (reprinted most recently in A. M. Turing, Collected works. Mechanical intelligence. Edited and with an introduction by \textit{D. C. Ince} [North-Holland Publishing Co., Amsterdam (1992; Zbl 0751.01018)]) which developed, in effect, machine models of connectionism, but which was apparently unknown to Rosenblatt and others working in this area. Turing described the principles of an ``unorganized machine'', a network of simple neuron-like elements. With each such unit operating under simple rules that determine one of its two possible states and rules for propagation and activation, the network could, Turing believed, be ``trained'' to perform a definite task starting in an initial random state. Also, he claimed that with sufficient units and appropriate initial conditions such an ``unorganized machine'' could be made into a universal Turing machine with given storage capacity. Turing did not attempt to give formal proofs of his claims and had not worked out the ``training'' algorithms later developed by Rosenblatt and by Clark and Farley in 1954. The authors give evidence of Turing's independence from U.S. developments in computers; even when he knew about such developments it is claimed that he took fundamentally different directions in his own work. In the same way, the authors maintain, he is likely not to have been influenced by the work on neural networks of McCulloch and Pitts in 1943 which had been influenced by Turing's own earlier work but which Turing may not have seen by 1948.